WO2023006205A1

WO2023006205A1 - Devices and methods for machine learning model transfer

Info

Publication number: WO2023006205A1
Application number: PCT/EP2021/071281
Authority: WO
Inventors: Dario BEGA; Abdelrahman ABDELKADER; Alberto Conte
Original assignee: Nokia Technologies Oy
Priority date: 2021-07-29
Filing date: 2021-07-29
Publication date: 2023-02-02

Abstract

A machine learning model consumer (102) configured to : - send a request for a first machine learning model to a machine learning model provider (103), the first machine learning model being trained and comprising at least two parts; - receive, from the machine learning model provider (103), a response on the request, the response comprising either a first part of the first machine learning model as an executable file or information to access a first part of the first machine learning model as a service, and metadata associated with the first part.

Description

DEVICES AND METHODS FOR MACHINE LEARNING MODEL TRANSFER TECHNICAL FIELD

[0001]. Various example embodiments related generally to devices and methods for machine learning model transfer.

BACKGROUND

[0002], Training of Machine Learning (ML) algorithms requires high computational resources to process a huge amount of training data required for performing the training task.

[0003]. In traditional data mining and machine learning algorithms, training of ML model is performed in an isolate way such that separate models are trained based on specific tasks and specific labeled or unlabeled training data to make predictions on future data.

[0004]. Transfer learning is a machine learning technique where a model (referred to as a ‘source model’) designed and trained for a task (referred to as a ‘source task’) in a first domain (referred to as a ‘source domain’) is used as starting point for building a new model (referred to as a ‘target model’) on a second task (referred to as a ‘target task’) in a second domain (referred to as ‘target domain’).

[0005]. The knowledge (comprising features and weights) from previously learned source tasks provided by the pre-trained source model can be leveraged for training newer target models. By taking the pre-trained source model as a starting point to train the target model, transfer learning enables reducing both the training time and the computational resources needed for training the target model and enables improving the performance of the target model on the target task. Further, in contrast to traditional machine learning, transfer learning allows the domains (source and target domains) and the tasks (source and target tasks) to be different. [0006]. Transfer learning is widely used on computer vision and natural language processing. In particular, transfer learning is widely applied in the context of deep learning in which deep learning systems and models are used as source models. The use of deep learning in transfer learning is also called deep transfer learning.

[0007]. Deep learning systems and models are layered architectures that learn different features at different layers. These layers are then finally connected to a last layer to get a final output. The layered architecture comprises a first set of layers (referred to as bottom layers) that store a global knowledge of the task statistics and a second set of layers (referred to as top layers) that store a local knowledge that is specific to the task.

[0008]. In order to take advantage of the global knowledge, only bottom layers of a deep learning model are used in deep transfer learning to build the target model. Accordingly, the weights of the bottom layers of the pre-trained source model are kept frozen (i.e., they will not be changed) and only the weights of the new untrained layers are updated during the training phase of the target model. After the training, the target model is specialized for the target task leveraging on the global knowledge provided by the pre-trained source model.

[0009]. Transfer learning is accordingly performed by transferring the whole trained source model as an executable. Such an approach is sub-optimal because it forces the ML model consumer to use the whole trained source model without modifications, which leads to lower performance compared to using a trained model for the specific task. And even if the ML model consumer is allowed to fine-tune the trained source model for the specific task using its own data, fine-tuning the trained source model would require the update of the whole source model, which requires a higher amount of computational resources.

[00010]. There is accordinlgy a need for enhanced reduced-complexity deep transfer learning techniques in particular and enhanced transfer learning in general.

SUMMARY

[00011]. The scope of protection is set out by the independent claims. The embodiments, examples and features, if any, described in this specification that do not fall under the scope of the protection are to be interpreted as examples useful for understanding the various embodiments or examples that fall under the scope of protection.

[00012]. In a first aspect, there is provided a machine learning model consumer configured to: [00013].- send a request for a first machine learning model to a machine learning model provider, the first machine learning model being trained and comprising at least two parts; [00014].- receive, from the machine learning model provider, a response on the request, the response comprising either a first part of the first machine learning model as an executable file or information to access a first part of the first machine learning model as a service, and metadata associated with the first part.

[00015]. In an embodiment when the received response comprises the first part of the first machine learning model as an executable file, the machine learning model consumer may be configured to generate a second machine learning model initially untrained and to generate a third machine learning model using the second machine learning model and the first part. [00016]. The machine learning model consumer may be configured to train the third machine learning model by training the second machine learning model using training data.

[00017]. The machine learning model consumer may be configured to process the third machine learning model using service data to produce data analytics.

[00018]. In an embodiment in which the second machine learning model is a deep learning network comprising a plurality of layers, the machine learning model consumer may be configured to generate the third machine learning model by chaining the plurality of layers to the first part. [00019]. In an embodiment when the received response comprises information to access a first part of the first machine learning model as a service, the machine learning model consumer may be configured to :

[00020].- generate a second machine learning model, the second machine learning model being initially untrained;

[00021].- send input training data to the machine learning model provider;

[00022].- receive output training data from the machine learning model provider, the output training data being generated by processing the first part using the input training data; and [00023].- use the output training data to train the second machine learning model.

[00024]. The machine learning model consumer may be configured to:

[00025].- send, to the machine learning model provider, input service data;

[00026].- receive, from the machine learning model provider, output service data, the output service data being generated by processing the first part using the input service data, and [00027].- use the output service data to process the trained second machine learning model to produce data analytics.

[00028]. In a second aspect, there is provided a machine learning model provider configured to:

[00029].- receive a request for a first machine learning model from a machine learning model consumer, the first machine learning model being trained and comprising at least two parts; [00030].- send, to the machine learning model consumer, a response on the request, the response comprising either a first part of the first machine learning model as an executable file or information to access a first part of the first machine learning model as a service, and metadata associated with the first part.

[00031]. The machine learning model provider may be configured to select the first machine learning model among one or more trained machine learning models depending on analytics service information comprised in the request for a first machine learning model.

[00032]. The request may comprise information to request only a first part of the first machine learning model.

[00033]. The first part may be trained to deliver output data from input data, the metadata associated with the first part comprising at least information indicating a format of the input data and a format of the output data.

[00034]. The metadata may comprise information related to the first part.

[00035]. The information related to the first part may comprise an identifier associated with the first part, a version of the first part, and information related to a machine learning task for which the first is trained.

[00036]. In a third aspect, there is provided a management data analytics service implementing the machine learning model consumer according to any preceding feature. [00037]. In a fourth aspect, there is provided a network data analytics function implementing the machine learning model consumer according to any preceding feature.

[00038], In a fifth aspect, there is provided a data structure for storage of metadata associated with a first part of a first machine learning model, the first part being trained to deliver output data from input data, the data structure comprising data elements related to the first part and data elements related to a format of the input data and a format of the output data.

[00039], The data elements related to the first part may comprise:

[00040].- a first data element representing an identifier associated with the first part ; [00041].- a second data element representing a version of the first part, and [00042].- a third data element representing information related to a learning task for which the first part is trained.

[00043], In a sixth aspect, there is provided a method comprising:

[00044].- sending a request for a first machine learning model to a machine learning model provider, the first machine learning model being trained and comprising at least two parts; [00045].- receiving from the machine learning model provider, a response on the request, the response comprising either a first part of the first machine learning model as an executable file or information to access a first part of the first machine learning model as a service, and metadata associated with said first part.

[00046], In a seventh aspect, there is provided a method comprising:

[00047], - receiving a request for a first machine learning model from a machine learning model consumer, the trained machine learning model being trained and comprising at least two parts; [00048].- sending to the machine learning model consumer a response on the request, the response comprising either a first part of the first machine learning model as an executable file or information to access a first part of the first machine learning model as a service, and metadata associated with the first part.

[00049], The method may comprise selecting the first machine learning model among one or more trained machine learning models depending on analytics service information comprised in the request for a first machine learning model.

[00050], In an eighth aspect, there is provided a program stored in a computer-readable non- transitory medium, the program comprising instructions stored on the computer-readable storage medium that, when executed by a processor, cause the processor to :

[00051].- send a request for a first machine learning model to a machine learning model provider, the first machine learning model being trained and comprising at least two parts; [00052].- receive, from the machine learning model provider, a response on the request, the response comprising either a first part of the first machine learning model as an executable file or information to access a first part of the first machine learning model as a service, and metadata associated with the first part. [00053]. In a nineth aspect, there is provided a program stored in a computer-readable non- transitory medium, the program comprising instructions stored on the computer-readable storage medium that, when executed by a processor, cause the processor to:

[00054].- receive a request for a first machine learning model from a machine learning model consumer, the first machine learning model being trained and comprising at least two parts; [00055].- send, to the machine learning model consumer, a response on the request, the response comprising either a first part of the first machine learning model as an executable file or information to access a first part of the first machine learning model as a service, and metadata associated with the first part.

[00056]. Advantageously, exemplary embodiments enable an ML model consumer to request only a part of a trained ML model in order to generate a new model in a same or different domain and/or for the same or different task.

[00057]. Advantageously, exemplary embodiments enable the sharing of an ML model as an executable or as a service.

[00058]. Advantageously, exemplary embodiments enable an ML model consumer to generate a new ML model with better performance on a new task, with a fast training process that requires lower computationanl resources.

[00059]. Advantageously, exemplary embodiments provide a machine learning service request that allows an ML model consumer to request access to only a trained part of a trained model.

[00060]. Advantageously, exemplary embodiments in applications to telecommunication networks enable the sharing of ML models in multi-vendor scenarios allowing to share partial functionality of a trained model without disclosing sensitive information.

[00061]. Further advantages will become clear to the skilled person upon examination of the drawings and the detailed description.

BREIEF DESCRIPTION OF THE DRAWINGS

[00062]. The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate example embodiments together with the general description given above, and the detailed description given below.

[00063]. FIG. 1 is a block diagram illustrating an exemplary network implementing a machine learning model consumer and a machine learning model provider, according to some embodiments.

[00064]. FIG. 2 is a connection flow illustrating the generation of data analytics using transfer learning, according to some embodiments. [00065]. FIG. 3 is a connection flow illustrating the generation of data analytics using transfer learning, according to some embodiments in which a first part of a trained ML model is provided as an executable to an ML model consumer.

[00066]. FIG. 4 is a connection flow illustrating the generation of data analytics using transfer learning, according to some embodiments in which a first part of a trained ML model is provided as a service to an ML model consumer.

[00067]. FIG. 5 is a block diagram illustrating a data structure for storing metadata associated with a first part of a trained ML model, according to exemplary embodiments.

[00068]. FIG. 6 is a block diagram illustrating an exemplary structure of a network entity operating in a telecommunication network, according to some embodiments.

[00069]. It should be noted that these figures are intended to illustrate the general characteristics of devices, methods, and structures utilized in certain example embodiments and to supplement the written description provided below. These drawings are not, however, to scale and may not precisely reflect the precise structural or performance characteristics of any given embodiment, and should not be interpreted as defining or limiting the range of values or properties encompassed by example embodiments. The use of similar or identical reference numbers in the various drawings is intended to indicate the presence of a similar or identical element or feature.

DETAILED DESCRIPTION

[00070]. Detailed example embodiments are disclosed herein. However, specific structural and functional details disclosed herein are merely representative for purposes of describing example embodiments. The example embodiments may, however, be embodied in many alternate forms and should not be construed as limited to only the embodiments set forth herein. Accordingly, while example embodiments are capable of various modifications and alternative forms, the embodiments are shown by way of example in the drawings and will be described herein in detail. It should be understood, however, that there is no intent to limit example embodiments to the particular forms disclosed.

[00071]. Specific details are provided in the following description to provide a thorough understanding of example embodiments. However, it will be understood by one of ordinary skill in the art that example embodiments may be practiced without these specific details. For example, systems may be shown in block diagrams so as not to obscure the example embodiments in unnecessary detail. In other instances, well-known processes, structures and techniques may be shown without unnecessary detail in order to avoid obscuring example embodiments.

[00072]. Exemplary embodiments provide devices, methods, and computer program products enabling enhanced transfer learning. [00073]. Exemplary transfer learning applications comprise, without limitation, natural language programming, computer vision (e.g. object recognition and identification), automatic speech recognition, classification (e.g. text classification, image classification, sentiment classification), clustering (e.g. text clustering, image clustering), reinforcement learning, collaborative filtering, sensor based location estimation, logical action model learning for Al planning and page ranking.

[00074]. Transfer learning involves two entitites : a machine learning model consumer (herein after referred to as an ‘ML model consumer’) and a machine learning model provider (herein after referred to as an ‘ML model provider’). The ML model consumer requests a trained ML model from the ML model provider to generate a new ML model. The ML model provider provides a trained ML model to the ML model consumer. The interaction between the ML model consumer and the ML model provider may use a request or subscription model and service-based interfaces.

[00075]. Exemplary embodiments provide devices, methods and computer program products enabling the use of transfer learning to generate data analytics in a telecommunication network.

[00076]. A telecommunication network is a system designed to transfer data from a network entity to one or more network entities. Data transfer involves data switching, transmission media, and system controls in addition to hardware and/or software resources that need to be deployed for data storage and/or processing.

[00077]. Data analytics enables converting input raw data generated by the network entities into information that can be processed, interpreted, and used for detailed analysis.

[00078]. Data analytics provision is defined according to a service-oriented approach described as an interaction between a data analytics consumer and a data analytics provider (that can be the data analytics producer). The data analytics consumer can request data analytics services or operations from the data analytics provider. The interaction between the data analytics consumer and the data analytics provider may use a request or subscription model and service-based interfaces.

[00079]. The generation of data analytics according to various embodiments rely on the use of a machine learning model that is generated based on transfer learning. In these embodiments, the ML model consumer is also the data analytics producer. The ML model consumer generates/produces data analytics, required by the data analytics consumer, using the trained model it acquired from the ML model provider via transfer learning.

[00080]. The following description of some embodiments will be made with reference to the use of transfer learning in the context of ML-based data analytics generation in a telecommunication network, for illustration purposes only. However, the skilled person will readily understand that the various embodiments may be used in other types of networks/systems for other applications of transfer learning involving an ML model consumer and an ML model provider, including the above mentioned applications of transfer learning. [00081]. FIG. 1 illustrates an exemplary telecommunication network 100 in which exemplary embodiments may be implemented. The telecommunication network 100 comprises an ML model consumer 102 configured to communicate with one or more data sources 104-i, with i varying from 1 to N, N being the total number of data sources.

[00082]. The ML model consumer 102 is also configured to communicate with an ML model provider 103 for the operations related to the transfer learning service, and to communicate with a data analytics consumer 101 for the operations related to the data analytics service. [00083]. The interaction between the ML model consumer 102 and the data sources 104-i, the ML model provider, and the data analytics consumer 101 may use a request or subscription model and service-based interfaces.

[00084]. The telecommunication network 100 may be a digital system part of a communication system, a data processing system, or a data storage system. Exemplary digital systems comprise, without limitations:

[00085]. -communication systems (e.g. radio communication systems, wireless communication systems, optical fiber-based communication systems, optical wireless communication systems, satellite communication systems);

[00086]. - storage systems (e.g. cloud computing systems);

[00087].- integer programming systems (e.g. computing systems, quantum computing systems);

[00088].- positioning systems (e.g. Global positioning systems or GPS, Global Navigation Satellite Systems or GNSS).

[00089]. According to some embodiments, the telecommunication network 100 may be: [00090].- wired (e.g. optical fiber-based networks);

[00091].- wireless (e.g. radio communication networks);

[00092]. - acoustic (e.g. underwater acoustic communication systems);

[00093]. - molecular (used for example in underground structures such as tunnels and pipelines).

[00094]. The network entities may be physically distributed over land, underwater, and in orbit. [00095]. In an exemplary application to wired data networks, the telecommunication network 100 may be a computer networking system in which one or more data sources 104-i are configured to operate in a wired network. Exemplary data sources 104-i adapted to such applications comprise computers, routers or switches connected to a small or large area wired network. Any type of physical cable may be used in such wired data network to ensure the transfer of data between the devices connected to the wired network comprising the one or more network data sources 104-i. [00096]. In another application to wireless networks, the telecommunication network 100 may be any wireless network involving any type of wireless propagation medium suitable for this type of connectivity. Exemplary wireless communication networks comprise, without limitation, ad-hoc wireless networks used in local area communications, wireless sensor networks, and radio communication networks (e.g. Long Term Evolution or LTE, LTE-advanced, 3G/4G/5G and beyond). In such applications, the one or more data sources 104-i may be any type of fixed or mobile wireless device/system/object configured to operate in a wireless environment. The one or more data sources 104-i may be remotely monitored and/or controlled. The one or more data sources 104-i may be equipped with one or more transmit antennas and one or more receive antennas.

[00097]. According to some embodiments in application to wireless networks, the data sources 104-i comprises, without limitations:

[00098].- user equipments (e.g. laptops, tablets, mobile phones, robots, Internet of Things (loT) devices);

[00099].- base stations (e.g. cellular base stations like eNodeB in LTE and LTE-advanced networks and gNodeB used in 5G networks, and femtocells used at homes or at business centers);

[000100]. - control stations (e.g. radio network controllers, base station controllers, network switching sub-systems);

[000101]. - network management systems responsible of the management and orchestration of the network elements operable in the telecommunication network 100.

[000102]. Exemplary applications to wireless networks comprise :

[000103]. - Machine-To-Machine (M2M) ;

[000104]. - Device-To-Device (D2D) ;

[000105]. - Industry 4.0 ;

[000106]. - Internet of Things or loT (for example vehicle-to-everything communications) involving networks of physical devices, machines, vehicles, home alliances and many other objects connected to each other and provided with a connectivity to the Internet and the ability to collect and exchange data without requiring human-to-human or human-to-computer interactions.

[000107]. In exemplary loT applications, the telecommunication network 100 may be a wireless loT network representing low energy power-consumption/long battery life/low- latency/low hardware and operating cost/high connection density constraints such as low- power wide area networks and low-power short-range loT networks. The telecommunication network 100 may be any wireless network enabling loT in licensed or license-free spectrum. [000108]. Exemplary wireless technologies used in loT applications may comprise: [000109]. - short range wireless networks (e.g. Bluetooth mesh networking, Light-Fidelity, Wi- FiTM, and Near-Field communications);

[000110]. - medium range wireless networks (e.g. LTE-advanced, Long Term Evolution- Narrow Band, NarrowBand loT), and

[000111]. - long range wireless networks (e.g. Low-Power Wide Area Networks (LPWANs), Very small aperture terminal, and long-range W-FiTM connectivity).

[000112]. Exemplary applications of M2M and loT applications comprise, without limitation: [000113]. - consumer applications (e.g. Internet of Vehicles, home automation, smart cities, wearable technologies, and connected health), and

[000114]. - commercial applications (e.g. digitalized healthcare connecting medical resources and healthcare services in which special monitors and sensors may be used to enable remote health monitoring and emergency notifications, smart traffic control, and road assistance). [000115]. According to some embodiments in application to wireless data networks, the one or more data sources 104-i are any physical system/device/object provided with the required hardware and/or software technologies enabling wireless communications and transfer of data or operational signals or messages to one or more network elements in the telecommunication network 100.

[000116]. In another application to optical fiber networks, the telecommunication network 100 may be any data network in which any optical fiber link is designed to carry data over short or long distances. Exemplary applications using optical fiber links over short distances comprise high-capacity networks such as data center interconnections. Exemplary applications using optical fiber links over long distances comprise terrestrial and transoceanic transmissions. In such applications, network data generated by the network elements operable in the telecommunication network 100 may be carried by optical signals polarized according to the different polarization states of the optical fiber. The optical signals propagate along the fiber- based link according to one or more propagation modes.

[000117]. Exemplary applications of optical fiber data networks comprise, without limitation, aerospace and avionics, data storage (e.g. in cloud computing systems, automotive, industry, and transportation). Such applications may involve transfer of voice (e.g. in telephony), data (e.g. data supply to homes and offices known as fiber to the home), images or video (e.g. transfer of internet traffic), or connection of networks (e.g. connection of switches or routers and data center connectivity in high-speed local area networks). In such applications, the one or more data sources 104-i may be any optical line terminal integrated for example at the provider’s central office or an optical network terminal deployed at the customer premises. [000118]. In another application to hybrid networks, the telecommunication network 100 may comprise wireless and optical connectivities between the network elements operable in the telecommunication network 100. For example, the telecommunication network 100 may be a hybrid optical-wireless access network in which a wireless base station sends data to a wireless gateway through an optical network unit. An exemplary architecture of a hybrid optical-wireless network comprises an integration of Ethernet Passive Optical Networks and wireless broadband communications based on WiMax (Worldwide Interoperability for Microwave Access) standardized in the IEEE 802.16 standards for access networks. In such applications, the one or more data sources 104-i may be any optical line terminal or optical network unit or any wireless device/system/sub-system/object.

[000119]. In another application to optical wireless data networks, connectivity between the network elements operable in the data network 100 may use optical communication in which unguided visible, infrared, or ultraviolet light is used to carry the signals carrying exchanged data (including network- related data and reports on network- related events). Exemplary optical wireless communications technologies comprise visible light communications (VLC), free space optics (FSO) and optical camera communications (OCC). Exemplary applications of optical wireless data networks comprise Optical Internet of Things supported by 5G networks. [000120]. The data sources 104-i are configured to provide data to the ML model consumer 102 for ML model training and data analytics generation. The data source 104-i, for i varying from 1 t N, may be any entity operable in the telecommunication network, providing and/or accessing to data that may be of any type.

[000121]. Exemplary types of data comprise, without limitation, management data, user data, subscription data, control data, network data, security data, activity data.

[000122]. Data used for training ML model is referred to as ‘training data’ and data used for generating data analytics is referred to as ‘service data’.

[000123]. Data may be of any nature, comprising for example, events, logs, performance measurements data, and local data analytics produced by or accessible to the data sources 104-i.

[000124]. The data analytics consumer 101 may require data analytics from the ML model consumer 102 to perform one or more actions that concern several domains comprising, without limitation, mobility management, session Management, Quality of Service (QoS) management, Application layer, security management, life cycle management, network performance management.

[000125]. For example, the data analytics consumer 101 may require data analytics for performing predictive analytics, anomaly detection, trend analysis, or clustering in a variety of use cases comprising :

[000126]. - network management ;

[000127]. - customer experience management ;

[000128]. - personalized marketing ;

[000129]. - load-level computation and prediction for a network slice instance ; [000130]. - Service experience computation and prediction for an application/User Equipment group ;

[000131]. - Network load performance computation and future load prediction ;

[000132]. - User Equipment expected behaviour prediction ;

[000133]. - User Equipment Abnormal behavior/anomaly detection ;

[000134]. - User Equipment Mobility-related information and prediction ;

[000135]. - User Equipment Communication pattern prediction ;

[000136]. - Congestion information - current and predicted for a specific location ;

[000137]. - QoS sustainability that involves reporting and predicting QoS change.

[000138]. In application to 5G and beyond telecommunication networks, a data source 104-i may be :

[000139]. - a User Equipment configured to send user data indicating for example the current status of the user equipment (e.g. battery, CPU, memory) ;

[000140]. - a Unified Data Repository configured for example to send subscription information ; [000141]. - a Network Management Service unit configured for example to send data related to the status of network and computing resources ;

[000142]. - an Access Network Discovery and Selection Function configured for example to send data related to available networks in a given area and related policies ;

[000143]. - an Online Charging System configured for example to send data related to credit management status ;

[000144]. - an Application Function configured for example to send data related to the user equipments accessing specific applications ;

[000145]. - a Policy Control Function configured for example to send data related to policies used at a given time.

[000146]. In an exemplary application to 5G and beyond telecommunication networks, the ML model consumer 102 may be implemented in a Network Data Analytics Function (NWDAF) defined in current 3GPP standards as a part of the 5G core network and used for performing data collection and providing network analytics information. In this application, data received from the data sources 104-i comprise network data. Data analytics generated by the ML model consumer in this case may be sent to one or more network functions and/or to an Operation, Administration and Management (OAM) entity. The data analytics consumer 101 in these cases may be an integrated part of the one or more network functions and/or the OAM. [000147]. In another exemplary application to 5G and beyong telecommunication netwoks, the ML model consumer 102 may be implemented in a Management Data Analytics Service (MDAS). The MDAS is defined in current 3GPP standards as a management entity configured to provide management data analytics to support network management and orchestration at the Radio Access Network level or at the Core Network level. In this application, data received from the data sources 104-i comprise management data.

[000148]. FIG. 2 is a connection flow illustrating the generation of data analytics using transfer learning, according to some embodiments.

[000149]. The data analytics consumer 101 triggers a data analytics service by sending, at step 200, a data analytics request, to the ML model consumer 102 (that represents the data analytics producer with respect to the data analytics consumer 101).

[000150]. Step 200 may be preceded by a discovery and selection phase (not illustrated in FIG. 2) during which the data analytics consumer selects a data analytics provider 102 that supports the requested analytics service and the required data analytics capabilities. [000151]. Discovery and selection procedures performed by the data analytics consumer 101 to select a data analytics provider may be performed using the methods defined in 3GPP standards and are not detailed in the present disclosure for a purpose of simplification. In the following description, the ML model consumer 102 corresponds to the data analytics provider selected by the data analytics consumer 101 during the discovery and selection phase. The discovery and selection procedures may comprise an authentication phase to authenticate the data analytics consumer 101.

[000152]. The data analytics request comprises analytics service information indicating specifications related to the required data analytics. For example, the analytics service information may comprise information indicating the name of analytics, an analytics identifier, the type of the required data analytics (e.g. statistics, predictions, notifications), and the scheduled time when the data analytics are needed.

[000153]. The ML model consumer 102 uses transfer learning to generate its own new ML model that is adapted to produce the data analytics according to the specifications required in the data analytics request. Transfer learning enables the ML model consumer 102 to speed up the generation and training of the new ML model.

[000154]. Transfer learning is thus triggered by the ML model consumer 102 by sending, at step 201 , a request for a first machine learning model to the ML model provider 103. The first machine learning model is a machine learning model that has been trained for a similar or different task, in a same or different domain. Similar task means that the ML model has been trained on similar data.

[000155]. The first machine learning model comprises at least two parts comprising a first part (also referred to as a ‘bottom part’) that has been trained to deliver output data from input data. [000156]. The first part is associated with metadata comprising information related to the first part and information indicating a format of the input data and a format of the output data. [000157]. For example, the information related to the first part comprises an identifier associated with the first part, a version of the first part, and information related to the task for which the first part is trained.

[000158]. In an embodiment using deep transfer learning, the first machine learning model is a deep learning model based on a multi-layer architecture (i.e. comprising a plurality of layers). In this embodiment, the first part comprises at least a part of the first layers of the trained deep learning model. In particular, the first part may correspond to the sub-model that comprises the first set of layers that store the global knowledge on the task on which the first part has been trained.

[000159]. Exemplary deep learning models comprise, without limitation, deep neural networks (e.g. autoencoders), deep belief networks, deep reinforcement learning, recurrent neural networks, and convolutional neural networks.

[000160]. Autoencoders comprise an input layer denoted by encoder and an output layer denoted by decoder.

[000161]. In embodiments using autoencoders, the first part corresponds to the encoder. [000162]. Step 201 may be preceded by a discovery and selection phase (not illustrated in FIG. 2) during which the ML model consumer 102 selects an ML model provider 103 that supports the requested ML model service and the required capabilities.

[000163]. Discovery and selection procedures performed by the ML model consumer 102 to select an ML model provider that supports the requested ML model service can be performed using the methods defined in 3GPP standards and are not detailed in the present disclosure for a purpose of simplification. In the following description, the ML model provider 103 corresponds to the ML model provider selected by the ML model consumer 102 during the discovery and selection phase. The discovery and selection phase may further comprise an authentication phase to authenticate the ML model consumer 102.

[000164]. In an embodiment, the request for a first machine learning model comprises information to request only the first part of the first machine learning model from the ML model provider 103.

[000165]. For example, in embodiments using an autoencoder as a trained deep learning model, the request comprises an indicator referred to as ‘Exdudejdecoder’ specifying that only the encoder of the autoencoder is required.

[000166]. In another embodiment, the request for a first machine learning model does not comprise information to request only a part of the machine learning model. In this embodiment, the ML model provider 103 decides to provide a part of the first machine learning model for example depending on the target task or the source task.

[000167]. At step 202, the ML model provider 103 selects a first machine learning model among one or more trained machine learning models. [000168]. In an embodiment in application to data analytics generation, the request for a first machine learning model comprises information extracted from the data analytics request to inform the ML model provider 103 about specifications of the use of the required trained first ML model for data analytics. In particular, the request for a first machine learning model comprises the analytics service information.

[000169]. In an exemplary embodiment, the ML model provider 103 selects, at step 202, the first machine learning model depending on the analytics service information.

[000170]. In an embodiment, the ML model provider 103 extracts, at step 202, the trained first part from the selected first machine learning model.

[000171]. The ML model provider 103 may extract the first part according to the request for the first machine learning model when the request specifies that only the first part is requested or may extract a part from the selected first machine learning model depending for example on the specifications of the use of the required first model, on the target task, or on the source task for which the first part of first ML model is trained.

[000172]. At step 203, the ML model provider 103 sends, to the ML model consumer 102, a response on the request for a first ML model, the response comprising the metadata associated with the first part.

[000173]. Depending on the availability and/or accessability of/to the first part, the response sent by the ML model provider 103 to the ML model consumer 102 comprises the first part as an executable file or comprises information to access the first part as a service.

[000174]. In the case of sending the first part as an executable file, the ML model consumer 102 cannot access the internals of the first part and does not have information on the architecture or parameters of the first part. In this case, the ML model provider 103 shares the first part of the trained first ML model without disclosing sensitive information.

[000175]. In the case of sending information to access the first part as a service, the ML model provider 103 does not share the first part, for example for security reasons, the trained first part can be consumed only as a service.

[000176]. At step 204, the ML model consumer 102 receives training data from one or more data sources 104-i.

[000177]. At step 205, the ML model consumer 102 generates a machine learning model (target model) using the training data received at step 204 for the model training and generation.

[000178]. At step 206, the ML model consumer 102 receives service data from one or more data sources 104-i.

[000179]. At step 207, the ML model consumer 102 processes the generated machine learning model using the received service data to produce data analytics. [000180]. At step 208, the ML model consumer 102 sends, to the data analytics consumer 101, a response on the data analytics request, the response comprising the produced data analytics.

[000181]. The generation of a machine learning model at step 205 using received training data and the prcessing of the generated machine learning model at step 207 using the received service data depend on whether the first part has been shared with the ML model consumer 102 as an executable file or as a service.

[000182]. FIG. 3 is a connection flow illustrating the generation of data analytics using transfer learning, according to some embodiments in which the first part is provided as an executable to the ML model consumer 102.

[000183]. Steps 300, 301 and 302 are respectively identical to steps 200, 201 and 202 described with reference to FIG. 2.

[000184]. At step 303, the ML model provider sends a response on the request for a first ML model, the response comprising the first part as an executable file and the metadata associated with the first part.

[000185]. Steps 304 to 307 are performed to generate and train an ML model using the first part received as an executable file, the metadata associated with the first part, and training data collected from one or more data sources 104-i.

[000186]. At step 304, the ML model consumer 102 generates a second machine learning model initially untrained and generates a third machine learning model using the received trained first part and the second untrained machine learning model. During this generation step, the ML model consumer 102 requires the information related to the formats of the input and output data of the trained first part.

[000187]. In an embodiment using deep transfer learning, the ML model consumer 102 generates the third ML model by chaining the layers of the second untrained machine learning model to the first part.

[000188]. The generated third ML model needs thus to be trained for the specific task (target task) required to provide the data analytics to the data analytics consumer 101. To do so, the ML model consumer 102 triggers a data collection service by sending, at step 305, a training data request to one or more data sources 104-i.

[000189]. At step 306, the ML model consumer 102 receives training data from the one or more data sources 104-i.

[000190]. At step 307, the ML model consumer 102 trains the third ML model generated at step 304 by training the second untrained machine learning model using the received traing data. The training phase is thus performed only over the new untrained second ML model, which is faster and requires lower computational resources compared with training the whole third ML model. [000191]. The generated and trained third ML model is thus ready for use to generate data analytics from service data.

[000192]. Then, at step 308, the ML model consumer 102 sends a service data request to one or more data sources 104-i.

[000193]. At step 309, the ML model consumer 102 receives service data from the one or more data sources 104-i.

[000194]. At step 310, the ML model consumer 102 processes the trained third ML model using the received service data to produce data analytics.

[000195]. At step 311, the ML model consumer 102 sends, to the data analytics consumer 101, a response on the data analytics request, the response comprising the produced data analytics.

[000196]. FIG. 4 is a connection flow illustrating the generation of data analytics using transfer learning, according to some embodiments in which the trained first part of the first ML model is provided as a service to the ML model consumer 102.

[000197]. Steps 400, 401 and 402 are respectively identical to steps 200, 201 and 202 described with reference to FIG. 2.

[000198]. At step 403, the ML model provider 103 sends a response on the request for a first ML model, the response comprising information to access the first part as a service and the metadata associated with the first part.

[000199]. In this case, the ML model consumer 102 needs to build its new model by generating an untrained model, then the ML model consumer 102 can use the first part service for training and processing the generated model. Since the first part is accessible as a service, the ML model consumer 102 can only send input data to the ML model provider 103 and exploit the output data obtained by processing the input data using the first part service.

[000200]. To do so, at step 404, the ML model consumer 102 generates an untrained model (referred to as a ‘second model) ML model. The generated second ML model is a new ML model that is initially untrained. The ML model consumer 102 uses the information related to the formats of the input and output data of the first part for generating the second ML model. [000201]. Training the second ML model for a specific task requires training data and access to the first part as a service.

[000202]. To do so, the ML model consumer 102 triggers a data collection service by sending, at step 405, a training data request to one or more data sources 104-i.

[000203]. At step 406, the ML model consumer 102 receives training data from the one or more data sources 104-i.

[000204]. At step 407, the ML model consumer 102 sends a first part service request to the ML provider 103, the first part service request comprising the training data as input training data and requesting output training data from the ML provider 103 such that the output training data are obtained by running the first part service on the input training data.

[000205], Upon receiving the first part service request, the ML model provider 103 runs, at step 408, the first part service using the received input training data to generate output training data.

[000206], At step 409, the ML model provider 103 sends, to the ML model consumer 102, a response on the first part service request, the response comprising the output training data generated by processing the first part using the input training data.

[000207], At step 410, the ML model consumer 102 trains the initially untrained second ML model generated at step 404 using as training data the output training data received from the ML model provider 103. The training phase is thus performed only over the new untrained second ML model, which is faster and requires low computational resources.

[000208], At step 411, the ML model consumer 102 sends a service data request to one or more data sources 104-i.

[000209], At step 412, the ML model consumer 102 receives service data from the one or more data sources 104-i.

[000210], In order to use the trained second ML model for generating data analytics from the received service data, the service data need to be processed by the first part service. [000211], Thus, at step 413, the ML model consumer sends a first part service request to the ML model provider 103, the first part service request comprising as input data the service data (also referred to as ‘input service data’) and requesting output data (also referred to as ‘output service data’) from the ML provider 103 such that the output data are obtained by processing the first part service using the input service data.

[000212], Upon receiving the first part service request, the ML model provider 103 runs, at step 414, the first part service using the received input service data to generate output service data.

[000213], At step 415, the ML model provider 103 sends, to the ML model consumer 102, a response on the first part service request, the response comprising the output service data generated by processing the first part using the input service data.

[000214], At step 416, the ML model consumer 102 processes the trained second ML model using the output service data to produce data analytics.

[000215], At step 417, the ML model consumer 102 sends, to the data analytics consumer 101, a response on the data analytics request, the response comprising the produced data analytics.

[000216], FIG. 5 is a block diagram illustrating an exemplary data structure for storing metadata associated with a first part of a trained first ML model, according to exemplary embodiments. [000217]. The data structure may be any data organization, management and storage format that enables access to and/or modification of stored data. Exemplary data structures comprise arrays, linked lists, records, and objects.

[000218]. The data structure comprises data elements. Each data element comprises an attribute field and a value field.

[000219]. The metadata associated with the first part comprise information related to the first part and information indicating a format of the input data and a format of the output data of the trained first part.

[000220]. For example, the information related to the trained first part comprises an identifier associated with the first part, a version of the first part, and information related to the task for which the first part is trained.

[000221]. Accordingly, the data structure 500 comprises data elements 501 related to the first part and data elements 502 related to a format of the input data and a format of the output data.

[000222]. For example, the data elements 501 comprise :

- a first data element 5010 representing an identifier associated with the trained first part. The attribute field comprised in the first data element 5010 may be ‘First part ID’ or ‘First part Identifier’ ;

- a second data element 5011 representing a version of the trained first part. The attribute field comprised in the second data element 5011 may be for example ‘First part Version’ ;

- a third data element 5012 representing information related to a learning task for which the first part is trained. The attribute field comprised in the third data element 5012 may be for example ‘Learning Task’.

[000223]. In an embodiment in application to data analytics generation, the third data element 5012 comprises the data analytics identifier comprised in the request for data analytics. [000224]. In an embodiment using deep transfer learning based on autoencoders, the attribute field comprised in the first data element 5010 is ‘Encoder ID’ and the attribute field comprised in the second data element 5011 is ‘Encoder Version’.

[000225]. The data elements 502 comprise a fourth data element 5020 representing information related to the input data format and a fifth data element 5021 representing information related to the output data format. The attribute field comprised in the fourth data element 5020 may be ‘Input data format’ and the attribute field comprised in the fifth data element 5021 may be Output data format’.

[000226]. 3GPP SA2 in TS 23.288 specifies requests on trained ML models at the NWDAF as an ML model consumer. According to these specifications, the NWDAF is allowed to request a trained model but cannot request a part of the trained model. Furthermore, ML model sharing is allowed only in a single-vendor scenario. [000227]. In the present disclosure, an adaptation of the existing ML model sharing service at the level of NWDAF is provided with the introduction of new attributes that enable the NWDAF requesting a part of the trained model and new attributes and metadata type that enable the NWDAF to receive a response indicating the sharing of the part of the trained model as an executable or as a service. The new services and attributes for deep transfer learning based on autoencoders comprise:

[000228], - a new attribute Exdudejdecoder’ for ‘Nnwdaf_MLModellnfo_Request’ service to allow the ML model consumer to specify that only the trained model encoder is required ; [000229], - a new service ‘Nnwdaf_MLModelAsaService_Request response’ to allow the ML model producer informing the ML model consumer that the request has been granted, and to provide the Meta Data about the trained encoder ;

[000230], - a novel service ‘Encoder_service’ to allow the ML model consumer sending input data to the trained encoder as a service ;

[000231], - a novel information Encoder output’ to allow the ML model producer to send to the ML model consumer the output generated by the trained encoder.

[000232], It should be appreciated by those skilled in the art that any functions, engines, block diagrams, flow diagrams, state transition diagrams and/or flowcharts herein represent conceptual views of illustrative circuitry embodying the principles of the invention. Similarly, it will be appreciated that any flow charts, flow diagrams, state transition diagrams, pseudo code, and the like represent various processes which may be substantially represented in computer readable medium and so executed by a computer or processing apparatus, whether such computer or processor is explicitly shown.

[000233], Each described computation function, block, step can be implemented in hardware, software, firmware, middleware, microcode, or any suitable combination thereof. If implemented in software, the computation functions, blocks of the block diagrams and/or flowchart illustrations can be implemented by computer program instructions / software code, which may be stored or transmitted over a computer-readable medium, or loaded onto a general purpose computer, special purpose computer or other programmable processing apparatus and / or system to produce a machine, such that the computer program instructions or software code which execute on the computer or other programmable processing apparatus, create the means for implementing the functions described herein.

[000234], When implemented in software, firmware, middleware or microcode, the program code or code segments to perform the necessary tasks may be stored in a machine or computer readable medium such as a computer readable storage medium. When implemented in software, a processor or processors will perform the necessary tasks. For example, as mentioned above, according to one or more example embodiments, at least one memory may include or store computer program code, and the at least one memory and the computer program code may be configured to, with at least one processor, cause an apparatus to perform the necessary tasks. Additionally, the processor, memory and example algorithms, encoded as computer program code, serve as means for providing or causing performance of operations discussed herein.

[000235], In the present description, block denoted as "means configured to” perform a certain function shall be understood as functional blocks comprising circuitry that is adapted for performing or configured to perform a certain function. A means being configured to perform a certain function does, hence, not imply that such means necessarily is performing said function (at a given time instant). Moreover, any entity described herein as "means", may correspond to or be implemented as "one or more modules", "one or more devices", "one or more units", etc. When provided by a processor, the functions may be provided by a single dedicated processor, by a single shared processor, or by a plurality of individual processors, some of which may be shared. Moreover, explicit use of the term "processor" or "controller" should not be construed to refer exclusively to hardware capable of executing software, and may implicitly include, without limitation, digital signal processor (DSP) hardware, network processor, application specific integrated circuit (ASIC), field programmable gate array (FPGA), read only memory (ROM) for storing software, random access memory (RAM), and non-volatile storage. Other hardware, conventional or custom, may also be included. Their function may be carried out through the operation of program logic, through dedicated logic, through the interaction of program control and dedicated logic, or even manually, the particular technique being selectable by the implementer as more specifically understood from the context.

[000236], As used in this application, the term “circuit” or “circuitry” may refer to one or more or all of the following:

[000237], (a) hardware-only circuit implementations (such as implementations in only analog and/or digital circuitry) and

[000238], (b) combinations of hardware circuits and software, such as (as applicable) : (i) a combination of analog and/or digital hardware circuit(s) with software/firmware and (ii) any portions of hardware processor(s) with software (including digital signal processor(s)), software, and memory(ies) that work together to cause an apparatus, such as a mobile phone or server, to perform various functions); and

[000239], (c) hardware circuit(s) and or processor(s), such as a microprocessor(s) or a portion of a microprocessor(s), that requires software (e.g., firmware) for operation, but the software may not be present when it is not needed for operation.”

[000240], This definition of “circuit” or “circuitry” applies to all uses of this term in this application, including in any claims. As a further example, as used in this application, the term circuitry also covers an implementation of merely a hardware circuit or processor (or multiple processors) or portion of a hardware circuit or processor and its (or their) accompanying software and/or firmware. The term circuitry also covers, for example and if applicable to the particular claim element, an integrated circuit for a network element or network node or any other computing device or network device. The term circuitry may cover digital signal processor (DSP) hardware, network processor, application specific integrated circuit (ASIC), field programmable gate array (FPGA), etc.

[000241]. The “circuit” or “circuitry” may be or include, for example, hardware, programmable logic, a programmable processor that executes software or firmware, and/or any combination thereof (e.g. a processor, control unit/entity, controller) to execute instructions or software and control transmission and receptions of signals, and a memory to store data and/or instructions. [000242], The “circuit” or “circuitry” may also make decisions or determinations, generate frames, packets or messages for transmission, decode received frames or messages for further processing, and other tasks or functions described herein. The circuitry may control transmission of signals or messages over a radio network, and may control the reception of signals or messages, etc., via a radio network (e.g., after being down-converted by radio transceiver, for example).

[000243], As disclosed herein, the term "storage medium," "computer readable storage medium" or "non-transitory computer readable storage medium” may represent one or more devices for storing data, including read only memory (ROM), random access memory (RAM), magnetic RAM, core memory, magnetic disk storage mediums, optical storage mediums, flash memory devices and/or other tangible machine-readable mediums for storing information. The term "computer-readable medium" may include, but is not limited to, portable or fixed storage devices, optical storage devices, and various other mediums capable of storing, containing or carrying instruction(s) and/or data.

[000244], The methods and devices described herein may be implemented by various means. For example, these techniques may be implemented in hardware, software, or a combination thereof. For a hardware implementation, the processing elements of the different network elements operating in the data network 100 can be implemented for example according to a hardware-only configuration (for example in one or more FPGA, ASIC, or VLSI integrated circuits with the corresponding memory) or according to a configuration using both VLSI and Digital Signal Processor (DSP).

[000245], FIG. 6 is a block diagram representing an exemplary hardware/software architecture of a network entity 600 operating in the telecommunication network 100 such as the network data sources 104-i, the data analytics consumer 101, the ML model consumer 102, and the ML model provider 103, according to some embodiments. As illustrated, the architecture may include various computing, processing, storage, communication, and displaying units comprising: - communication circuitry comprising a transceiver 602 (e.g. wireless or optical transceiver) configured to connect the network entity 600 to corresponding links in the telecommunication network 100, and to ensure transmission/reception of data and/or signals. The communication circuitry may support various network and air interface such as wired, optical fiber, and wireless networks;

- a processing unit 603 configured to execute the computer-executable instructions to run the methods and algorithms according to the various embodiments and perform the various required functions of the network entity such as data analytics production, ML models training and processing, training data processing, service data processing (e.g. input/output processing) and any functionalities required to enable the network entity 600 to operate in the telecommunication network 100 according to the various embodiments. The processing unit 602 may be a general purpose processor, a special purpose processor, a DSP, a plurality of microprocessors, a controller, a microcontroller, an ASIC, an FPGA circuit, any type of integrated circuit, and the like;

- a power source 604 that may be any suitable device providing power to the network entity 600 such as dry cell batteries, solar cells, and fuel cells;

- a localization unit 605 such as a GPS chipset implemented in applications that require information indicating the location of the network entity 800;

- a storage unit 606 possibly comprising a random access memory (RAM) ora read-only memory used to store processed data (e.g. network-related data, reports on network- related events) and any data required to perform the functionalities of the network entity 600 according to the embodiments;

- Input peripherals 607;

- Output peripherals 608 comprising communication means such as displays enabling for example man-to-machine interaction between the network entity 600 and the telecommunication network 100 administrator for example for configuration and/or maintenance purposes. [000246], The architecture of the device 600 may further comprise one or more software and/or hardware units configured to provide additional features, functionalities and/or network connectivity.

[000247], Furthermore, the methods described herein can be implemented by computer program instructions supplied to the processor of any type of computer to produce a machine with a processor that executes the instructions to implement the functions/acts specified herein. These computer program instructions may also be stored in a computer-readable medium that can direct a computer to function in a particular manner. To that end, the computer program instructions may be loaded onto a computer to cause the performance of a series of operational steps and thereby produce a computer implemented process such that the executed instructions provide processes for implementing the functions specified herein. [000248], For example, the program comprises instructions stored on the computer-readable storage medium that, when executed by a processor, cause the processor to:

[000249], - send a request for a first machine learning model to a machine learning model provider, the first machine learning model being traind and comprising at least two parts; [000250], - receive, from the machine learning model provider, a response on the request, the response comprising either a first part of the first machine learning model as an executable file or information to access a first part of the first machine learning model as a service, and metadata associated with the first part.

[000251], In another example, the program comprises instructions stored on the computer- readable storage medium that, when executed by a processor, cause the processor to:

[000252], - receive a request for a first machine learning model from a machine learning model consumer, the first machine learning model being trained and comprising at least two parts; [000253], - send, to the machine learning model consumer, a response on the request, the response comprising either a first part of the first machine learning model as an executable file or information to access a first part of the first machine learning model as a service, and metadata associated with the first part.

[000254], It will be understood that the configurations and/or approaches described herein are exemplary in nature, and that these specific embodiments or examples are not to be considered in a limiting sense, because numerous variations are possible. The specific routines or methods described herein may represent one or more of any number of processing strategies. As such, various acts illustrated and/or described may be performed in the sequence illustrated and/or described, in other sequences, in parallel, or omitted. Likewise, the order of the above-described processes may be changed.

[000255], The subject matter of the present disclosure includes all novel and non-obvious combinations and sub-combinations of the various processes, systems and configurations, and other features, functions, acts, and/or properties disclosed herein, as well as any and all equivalents thereof.

Claims

1. A machine learning model consumer (102) configured to : send a request for a first machine learning model to a machine learning model provider (103), the first machine learning model being trained and comprising at least two parts; receive, from the machine learning model provider (103), a response on the request, the response comprising either a first part of the first machine learning model as an executable file or information to access a first part of the first machine learning model as a service, and metadata associated with said first part.

2. The machine learning model consumer (102) of claim 1, wherein when the received response comprises the first part of the first machine learning model as an executable file, the machine learning model consumer (102) is configured to generate a second machine learning model initially untrained and to generate a third machine learning model using the second machine learning model and said first part.

3. The machine learning model consumer (102) of claim 2, wherein the machine learning model consumer (102) is configured to train the third machine learning model by training the second machine learning model using training data.

4. The machine learning model consumer (102) of any preceding claim 2 or 3, wherein the machine learning model consumer (102) is configured to process the third machine learning model using service data to produce data analytics.

5. The machine learning model consumer (102) of any preceding claim 2 to 4, wherein the second machine learning model is a deep learning network comprising a plurality of layers, the machine learning model consumer (102) being configured to generate the third machine learning model by chaining the plurality of layers to said first part.

6. The machine learning model consumer (102) of claim 1, wherein when the received response comprises information to access a first part of the first machine learning model as a service, the machine learning model consumer (102) is configured to :

- generate a second machine learning model, the second machine learning model being initially untrained;

- send input training data to the machine learning model provider (103) ;

- receive output training data from the machine learning model provider (103), the output training data being generated by processing said first part using the input training data ; and

- use the output training data to train the second machine learning model.

7. The machine learning model consumer (102) of claim 6, wherein the machine learning model consumer (102) is configured to : - send, to the machine learning model provider (103), input service data ;

- receive, from the machine learning model provider (103), output service data, the output service data being generated by processing said first part using the input service data, and

- use the output service data to process the trained second machine learning model to produce data analytics.

8. A machine learning model provider (103) configured to : receive a request for a first machine learning model from a machine learning model consumer (102), the first machine learning model being trained and comprising at least two parts; send, to the machine learning model consumer (102), a response on the request, the response comprising either a first part of the first machine learning model as an executable file or information to access a first part of the first machine learning model as a service, and metadata associated with said first part.

9. The machine learning model provider (103) of claim 8, wherein the request comprises analytics service information, the machine learning model provider (103) being configured to select the first machine learning model among one or more trained machine learning models depending on the analytics service information.

10. The machine learning model consumer (102) of any preceding claim 1 to 7 or the machine learning model provider (103) according to any preceding claim 8 or 9, wherein the request comprises information to request only a first part of the first machine learning model.

11. The machine learning model consumer (102) of any preceding claim 1 to 7 and 10 or the machine learning model provider (103) of any preceding claim 8 to 10, wherein said first part is trained to deliver output data from input data, the metadata associated with said first part comprising at least information indicating a format of the input data and a format of the output data.

12. The machine learning model consumer (102) of any preceding claims 1 to 7 and 10 to 11 or the machine learning model provider (103) of any preceding claim 8 to 11, wherein the metadata comprises information related to the first part.

13. The machine learning model consumer (102) of claim 12 or the machine learning model provider (103) of claim 12, wherein the information related to the first part comprises an identifier associated with the first part, a version of the first part, and information related to a machine learning task for which the first is trained.

14. A management data analytics service implementing the machine learning model consumer (102) of any preceding claim 1 to 7 and 10 to 13.

15. A network data analytics function implementing the machine learning model consumer (102) of any preceding claim 1 to 7 and 10 to 13.

16. Data structure (500) for storage of metadata associated with a first part of a first machine learning model, said first part being trained to deliver output data from input data, wherein the data structure comprises data elements (501) related to the first part and data elements (502) related to a format of the input data and a format of the output data.

17. The data structure of claim 16, wherein the data elements (501) related to the first part comprise: a first data element (5010) representing an identifier associated with the first part ; a second data element (5011) representing a version of the first part, and a third data element (5012) representing information related to a learning task for which the first part is trained.

18. A method comprising : sending (201) a request for a first machine learning model to a machine learning model provider, the first machine learning model being trained and comprising at least two parts; receiving (203) from the machine learning model provider, a response on the request, the response comprising either a first part of the first machine learning model as an executable file or information to access a first part of the first machine learning model as a service, and metadata associated with said first part.

19. A method comprising : receiving (201 ) a request for a first machine learning model from a machine learning model consumer, the trained machine learning model being trained and comprising at least two parts; sending (203) to the machine learning model consumer a response on the request, the response comprising either a first part of the first machine learning model as an executable file or information to access a first part of the first machine learning model as a service, and metadata associated with said first part.

20. The method of claim 19, wherein the request for a first machine learning model comprises analytics service information, the method comprising selecting (202) the first machine learning model among one or more trained machine learning models depending on the analytics service information.

21. A program stored in a computer-readable non-transitory medium, the program comprising instructions stored on the computer-readable storage medium that, when executed by a processor, cause the processor to : send a request for a first machine learning model to a machine learning model provider, the first machine learning model being trained and comprising at least two parts; receive, from the machine learning model provider, a response on the request, the response comprising either a first part of the first machine learning model as an executable file or information to access a first part of the first machine learning model as a service, and metadata associated with the first part.

22. A program stored in a computer-readable non-transitory medium, the program comprising instructions stored on the computer-readable storage medium that, when executed by a processor, cause the processor to : receive a request for a first machine learning model from a machine learning model consumer, the first machine learning model being trained and comprising at least two parts; send, to the machine learning model consumer, a response on the request, the response comprising either a first part of the first machine learning model as an executable file or information to access a first part of the first machine learning model as a service, and metadata associated with the first part.